Innovative Prompting Strategies and Holistic Evaluation of LLM Movie Recommender

Recommendation systems play a crucial role in to-day's digital environment, and the rise of Large Language Models (LLMs) has created new opportunities to enhance their effectiveness. LLMs have shown great promise in improving recommendation systems by better understanding user preferences and generating personalized suggestions. However, zero-shot in-context learning often yields suboptimal performance, while fine-tuning requires significant resources. This study creatively develops four distinct prompting strategies: Basic, Context-based, Personalized, and Chain-of-Thought. Each strategy takes advantage of LLMs for movie recommendations, integrating collaborative filtering techniques with reasoning methods. Our innovative approach outperforms traditional deep learning methods without the need for fine-tuning. Unlike previous research that focused solely on recommendation accuracy, we use a comprehensive evaluation framework that includes accuracy, diversity, and novelty metrics. This holistic analysis highlights the specific strengths of each prompting strategy across different LLM architectures, providing valuable insights into balancing recommendation quality with user experience.